AutoDO: Robust AutoAugment for Biased Data with Label Noise via Scalable Probabilistic Implicit Differentiation

CVPR 2021 preprint: https://arxiv.org/abs/2103.05863

Abstract

AutoAugment has sparked an interest in automated augmentation methods for deep learning models. These methods estimate image transformation policies for train data that improve generalization to test data. While recent papers evolved in the direction of decreasing policy search complexity, we show that those methods are not robust when applied to biased and noisy data. To overcome these limitations, we reformulate AutoAugment as a generalized automated dataset optimization (AutoDO) task that minimizes the distribution shift between test data and distorted train dataset. In our AutoDO model, we explicitly estimate a set of per-point hyperparameters to flexibly change distribution of train data. In particular, we include hyperparameters for augmentation, loss weights, and soft-labels that are jointly estimated using implicit differentiation. We develop a theoretical probabilistic interpretation of this framework using Fisher information and show that its complexity scales linearly with the dataset size. Our experiments on SVHN, CIFAR-10/100, and ImageNet classification show up to 9.3% improvement for biased datasets with label noise compared to prior methods and, importantly, up to 36.6% gain for underrepresented SVHN classes.

BibTex Citation

If you like our paper or code, please cite it using the following BibTex:

@InProceedings{Gudovskiy_2021_CVPR,
    author    = {Gudovskiy, Denis and Rigazio, Luca and Ishizaka, Shun and Kozuka, Kazuki and Tsukizawa, Sotaro},
    title     = {AutoDO: Robust AutoAugment for Biased Data With Label Noise via Scalable Probabilistic Implicit Differentiation},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2021},
    pages     = {16601-16610}
}

Installation

Clone this repository: tested on Python 3.6
Install PyTorch: tested on v1.6
Install Kornia: tested on v0.4.1
Other dependencies in requirements.txt

Datasets

Model checkpoints, datasets and index files for distorted train data are saved by default into ./local_data/{dataset} folder. For example, MNIST data is saved into ./local_data/MNIST folder. In order to get statistically significant results, we execute multiple runs of the same configuration with randomized weights and training dataset splits and save results to ./local_data/{dataset}/runN/ folders. We suggest to check that you have enough disk space for large-scale datasets.

MNIST, SVHN, SVHN_extra, CIFAR-10, CIFAR-100

Datasets will be automatically downloaded

ImageNet

Due to large size, ImageNet has to be manually downloaded according to torchvision instructions.

Code Organization

./custom_datasets - contains dataloaders (copied from torchvision)
./custom_models - contains CNN architectures and AutoDO models and hyperparameter optimization function in automodels.py
./custom_transforms - contains policy models for RandAugment, Fast AutoAugment and DADA methods as well as common image preprocessing functions

Running Experiments

Install minimal required packages using requirements.txt
Run code with the config by selecting IR, NR, dataset and runN
The sequence below should reproduce SVHN reference results for run0:

python3 -m pip install -U -r requirements.txt
python3 implicit-augment.py -r run0 --gpu 0 -nr 0.1 -ir 100 --dataset SVHN --aug-model NONE
python3 implicit-augment.py -r run0 --gpu 0 -nr 0.1 -ir 100 --dataset SVHN --aug-model RAND
python3 implicit-augment.py -r run0 --gpu 0 -nr 0.1 -ir 100 --dataset SVHN --aug-model AUTO
python3 implicit-augment.py -r run0 --gpu 0 -nr 0.1 -ir 100 --dataset SVHN --aug-model DADA
python3 implicit-augment.py -r run0 --gpu 0 -nr 0.1 -ir 100 --dataset SVHN --aug-model SHA --los-model NONE --hyper-opt HES
python3 implicit-augment.py -r run0 --gpu 0 -nr 0.1 -ir 100 --dataset SVHN --aug-model SEP --los-model NONE --hyper-opt HES
python3 implicit-augment.py -r run0 --gpu 0 -nr 0.1 -ir 100 --dataset SVHN --aug-model SEP --los-model WGHT --hyper-opt HES
python3 implicit-augment.py -r run0 --gpu 0 -nr 0.1 -ir 100 --dataset SVHN --aug-model SEP --los-model BOTH --hyper-opt HES

After finishing all N runs, we calculate [mu/std] results using the following script:

python3 get-results.py -nr 0.1 -ir 100 --dataset SVHN --aug-model NONE
python3 get-results.py -nr 0.1 -ir 100 --dataset SVHN --aug-model RAND
python3 get-results.py -nr 0.1 -ir 100 --dataset SVHN --aug-model AUTO
python3 get-results.py -nr 0.1 -ir 100 --dataset SVHN --aug-model DADA
python3 get-results.py -nr 0.1 -ir 100 --dataset SVHN --aug-model SHA --los-model NONE --hyper-opt HES
python3 get-results.py -nr 0.1 -ir 100 --dataset SVHN --aug-model SEP --los-model NONE --hyper-opt HES
python3 get-results.py -nr 0.1 -ir 100 --dataset SVHN --aug-model SEP --los-model WGHT --hyper-opt HES
python3 get-results.py -nr 0.1 -ir 100 --dataset SVHN --aug-model SEP --los-model BOTH --hyper-opt HES

Ablations studies are performed using "--overfit/oversplit" arguments
Learning curves are logged by tensorboard inside implicit-augment.py script
Qualitative Figures are generates using visualize-density.py script

gudovskiy / autodo